Anthropic says it has safeguards in place to prevent common risks like prompt injection, and it will limit access to certain “off limits” apps (e.g., “investment and trading platforms, cryptocurrency”) by default.
Anthropic also notes on a support page that the model is trained to avoid “risky operations” such as moving or investing money, modifying files, scraping facial images, or inputting “sensitive data.” But the company also warns that such training safeguards “aren’t perfect” and “aren’t absolute,” meaning that “Claude may occasionally act outside these boundaries.”
Anthropic also warns that, when computer use is activated, Claude will be able to see anything visible on-screen, including “personal data, sensitive documents, or private information.” For all these reasons, the company recommends “starting with the apps you trust and not working with sensitive data” during this research preview stage.
Anthropic’s announcement comes just weeks after the rollout of Perplexity’s Personal Computer, Manus’s My Computer, and Nvidia’s NemoClaw, which similarly let their AI agents take direct control of the desktop. All of these corporate moves follow the viral spread of OpenClaw earlier in the year, which led OpenAI to hire OpenClaw creator Peter SteinBerger “to drive the next generation of personal agents.”
Source: arstechnica.com
